Aiming at the need to use vehicle key points to obtain vehicle 3D posture in smart transportation and self-driving systems, a vehicle key point detection model based on CPN network is proposed. The model integrates deep semantic information and shallow spatial resolution information in the form of a U-type structure with ResNet50 as the backbone network to build a Gaussian heat map pyramid. Then, SoftArgmax is used to decode the key point coordinates from the Gaussian heat map end to the end. The vehicle key point detection model is trained on a training set of 200000 sheets, which can predict the coordinates and visibility of 78 key points on defined cars and SUV models at the same time. The normalized pixel error of the prediction point under the input image of 256×256 is 1.57, and the visibility prediction of the point reaches the accuracy of 0.96 at the recall rate of 0.95. The experimental results show that the vehicle key point detection model based on the CPN network has high accuracy and has been applied to intelligent transportation systems in Beijing, Wuhan and other cities.
